GPT-6 Cyber Explained: Critical Rating, Price and Access

OpenAI released GPT-6 Cyber on 5 November 2026, a security-focused model built to find software vulnerabilities and develop working exploits. It is not a model you can sign up for: access is limited to approved customers. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, the first model of its cyber line to reach that level.

Quick summary

• What it is: a variant of GPT-6 Astra with extra security training.
• Access: API and Codex, initially for a select group of approved Daybreak Red customers, with wider access promised.
• Price: $20 per million input tokens and $100 per million output tokens (VentureBeat, from OpenAI's docs).
• Headline score: 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber.
• Safety rating: Critical, up from High.

What can GPT-6 Cyber do?

OpenAI says it is much better at turning a known weakness into a working exploit and at staying inside the limits a researcher sets. On ExploitGym it scored 48.5%, about 16 points above its predecessor. In a web penetration test without source code access it found 40.3% of vulnerabilities, against 32% for GPT-6 Astra. On a new RemoteExploitBench covering GitLab, PostgreSQL and Jenkins, OpenAI says it completed 95% of tasks while Astra, running under the more restricted Daybreak Blue setup, refused all of them.

What did it find in Linux and Firefox?

OpenAI says its researchers reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948, compared with 166 for GPT-5.6 Sol on the same setup. It disclosed 304 critical findings to maintainers with candidate patches. In Firefox, it produced evidence for 81 previously unknown vulnerabilities that Mozilla classed as high severity, and Mozilla has started shipping fixes. VentureBeat cautions that the 948 exploit count should not be read as 948 new, unpatched bugs.

Why does staying in scope matter?

A key risk with security agents is wandering outside the assigned target. In an ExploitGym Honeypot test, where easy-to-attack side services tempt the model, GPT-6 Cyber stayed in scope 99.93% of the time without a deployment prompt (94.49% for GPT-5.6 Cyber) and 100% with one. OpenAI links that test to the earlier Hugging Face incident. These are results on a benchmark, not guarantees about real use.

How does the price compare with rivals?

Prices are standard API rates as compiled by VentureBeat on 5 November. Token price is not cost per finding, since models use different amounts of context to reach a validated result. For the wider model race see our AI coding comparison and the Gemini 4 Carbon explainer.

Can regular users or Indian developers use it?

Not right now. Individuals and organisations can apply to the Daybreak program and upload details of their defensive use, and OpenAI says it will expand Daybreak Red access soon. It also runs a $1 billion Daybreak for Frontline Defenders grant program for resource-constrained teams. Everyday ChatGPT users are not affected.

What are the safeguards against misuse?

OpenAI pairs the model with cyber-specific misalignment monitoring, targeted monitoring for emerging threats (it cites the coordinated disclosure of critical PHP flaws), and Private Safety Processing, which reviews encrypted records kept in the customer's own storage for 30 days for API customers on Zero Data Retention. Daybreak Red API access requires IP allowlisting by default, to limit damage if credentials are stolen. OpenAI also recommends sandboxing, action review and human oversight for high-risk workflows.

What about critical infrastructure?

A new Daybreak Red for Critical Infrastructure group lets approved defenders test power, water and industrial control systems under safeguards. OpenAI worked with industrial security firm Dragos, which reported that a frontier model on a mock power plant range separated operational technology assets from IT and found two zero-days in engineering software. Dragos warned that the skill needed for reconnaissance before an attack is falling, which is why OpenAI says it tightened blocks on help that could compromise safety-critical systems.

Verdict

The numbers are OpenAI's own and mostly come from its own benchmarks, so independent testing will matter. Still, the Firefox and Linux results are concrete, and the Critical rating signals how seriously the company treats the dual-use risk.

Related articles

GPT-6.1 Terra: Price, Context and How It Compares to SolClaude Personal Finance: Plaid Bank Access ExplainedChatGPT Images 2.6 Adds GIF Creation and Two New API ModelsSarvam Prism Explained: India's 44 ms Open Decision Model